NiubiGEO has published an Apache-2.0 workbench for testing how AI models describe a product, which competitors they mention and what sources they return. Developers can inspect the answers and execution conditions behind a report instead of relying only on a visibility score, according to the project repository.
A NiubiGEO project starts with a domain. Users select one or more OpenRouter models and configure web search separately for each model before running a test. Models answer independently, and a failure from one model does not remove successful results from the others; failed models can also be retried separately. The saved results include the actual execution conditions for each run, as described in the repository documentation.
The initial domain test reports product descriptions, categories, competitors and associated keywords. Users can then confirm keywords and test them without naming the target brand. NiubiGEO’s methodology distinguishes recognition after a direct brand prompt from an unprompted appearance, and it treats a mention, a positive description and an explicit recommendation as different findings. Those distinctions are documented in the project’s examples and usage notes.
NiubiGEO displays original answers, provider citations and ordinary URLs separately while retaining failures and unresolved findings. This evidence can help users locate a statement or source, but the project cautions that a citation alone does not explain why a model recommended a product. Its documentation also says that publishing an article does not guarantee an AI recommendation.
Repeated measurements and scheduled monitoring create a history tied to the underlying answers. Earlier records remain available when the selected models change, while scheduled execution requires the monitoring worker to be running. The project explicitly warns that a few closely spaced tests demonstrate repeated testing rather than long-term growth. The monitoring notes do not present any single answer as a permanent ranking.
The local quick start requires Node.js 22 or later and the user’s own OpenRouter API key. Published cases can be explored without installation or an API key, but testing another product incurs model and search API charges; operators also cover their hosting costs. NiubiGEO separately advertises paid AI testing by real people and GEO optimization services through the repository.
The project positions self-hosting, model choice and access to source evidence as its distinguishing criteria. It suggests considering commercial platforms when hosted services, marketing workflows or an existing search dataset take priority, while stating that its vendor comparisons are selection suggestions rather than a controlled benchmark or ranking. That limitation appears in the comparison guidance.
Analysis: NiubiGEO’s most defensible value is auditability, not proof of a stable level of AI visibility. A stored response can establish what a selected model returned under recorded conditions, giving developers a concrete artifact to inspect when a product is omitted or described inaccurately. It cannot, by itself, establish why the result occurred or whether it will persist—limits consistent with the project’s warnings about recommendations, citations and short-term measurements in the repository.
The unresolved trade-off is breadth versus comparability. Adding models, search modes and scheduled runs expands the set of observations, but it also adds model and search charges and creates more execution conditions to compare. NiubiGEO records those conditions; it does not claim that doing so converts variable model answers into a permanent ranking. The project documentation instead frames the records as evidence for further investigation.